Deep Dive
What Claude Skills Are and Why They Matter
Eliot opens by positioning himself as someone who has built over 450 Claude skills through a year of daily work for his business, students, and consulting clients. A Claude skill is fundamentally a set of step-by-step instructions stored in markdown files that Claude keeps and executes on command. The core innovation isn't that you write better individual prompts — it's that you capture a repeatable process once, store it, and then never have to copy-paste or re-explain it again. Skills can fire automatically when Claude detects you're asking for that type of work, or you can explicitly invoke them via slash commands. The benefit is consistency: every time Claude runs the skill, it follows the same process and produces the same quality of output. Unlike storing a great prompt in a document, skills are integrated into Claude's workflow, meaning they can run without you consciously remembering to activate them. This solves the friction problem of searching for old prompts and editing them each time.
The Anatomy of a Claude Skill: Three Core Components
Every Claude skill has three main parts. First is the description, which explains what the skill does and includes trigger phrases — the natural language cues that tell Claude when to fire the skill automatically. Second is the skill.md file, the core instruction set that contains the step-by-step workflow, decision logic, and the order in which things happen. This is where you encode how Claude should think about the problem and what it should never do. Third are reference files, which store supporting detail like templates, checklists, style guides, worked examples, and even Python scripts for automation. The distinction matters: skill.md holds the decisions Claude needs to make along the way, while reference files hold the lookup tables and supporting information it references when needed. For example, in an SEO blog writer skill, the main file says 'apply my full style guide at this stage,' and the reference file contains the actual multi-page style guide. Trigger phrases are critical because they determine when a skill fires automatically — vague phrases like 'write this transcript' might accidentally fire when you don't intend it, while specific ones like 'turn this YouTube transcript into an SEO blog post' activate only when truly relevant.
Two Types of Skills: Deterministic and Non-Deterministic
Claude skills fall into two categories based on their output characteristics. Deterministic skills always produce the same templated output — think invoices, audits, monthly reports, or structured documents where you're essentially swapping in new data into a fixed format. These are easy to build and easy to debug because failures are loud and obvious: if an invoice is missing a column or a template element is wrong, you immediately see it. Non-deterministic skills apply judgment, taste, and voice across varied situations, like brand voice rules or thinking frameworks that you apply to many different contexts. These fail quietly — they might look good the first run but gradually drift from your standards without a clear signal. They require more iterative training and closer monitoring. The SEO blog writer skill is non-deterministic because it carries judgment about keyword placement, tone, and structure across different topics. Understanding which type you're building helps you set realistic expectations for build time and maintenance overhead. Deterministic skills tend to be faster and more stable; non-deterministic ones are more flexible but need ongoing refinement.
Five Levels of Skill Complexity and Orchestration
Skills progress through five levels of complexity. Level 1 is a simple wrapped prompt that runs automatically with no extra input needed — fire and forget. Level 2 requires more information from you; it asks questions before executing so you give it the context it needs. Levels 1 and 2 are where most people stay because they feel like normal prompt engineering. Level 3 connects the skill to outside tools via connectors — your skill can pull data from Gmail, YouTube, Google Search Console, or push blog posts to Notion or publish them to your website. Level 4 chains skills together so they hand off work to each other in sequence: an audit feeds into a strategy skill, which feeds into a content creation skill, which feeds into a publishing skill. Level 5 adds an orchestrator or manager skill that sits on top of the entire team, understanding what each skill does and when to invoke them. You can tell the orchestrator 'my website traffic is down' and it knows which skills to call and how to chain them together. Critically, Eliot emphasizes that levels are not a scoreboard — a Level 1 skill that frees two hours weekly is a bigger win than a Level 5 system requiring constant troubleshooting and maintenance. The goal is time and money saved, not technical complexity for its own sake.
Method 1: Wrapping an Existing Prompt Into a Skill
The simplest build method takes a prompt you already use and wraps it into a skill via the skill creator skill. First, you enable the skill creator skill in your account — it should be on by default. Then you invoke it by typing forward slash and searching for 'skill creator' or simply saying 'build a Claude skill' in conversation. You paste in your existing prompt and ask Claude to package it into a skill. The skill creator skill has studied Anthropic's technical requirements and knows how to structure the skill.md, add proper trigger phrases, and bundle any reference files. Within minutes, Claude analyzes your prompt, builds the skill, and suggests a name and trigger phrases. You review the output, make any tweaks, and click save. This method works best when you already have a dialed prompt you've spent time refining — you're not starting from scratch, just moving a proven tool into the skill system. Eliot demonstrates this with an SEO audit skill: he pastes in a detailed prompt about auditing website visibility, Claude wraps it, and within minutes he can run it on any website by saying 'run an SEO audit on [URL].' The skill immediately produces a structured report with findings, fix lists, strategy recommendations, and evidence citations.
Method 2: Building from Cold Start Through Iterative Testing
When you don't have an existing prompt, you build from cold start by working through the process with Claude, testing, refining, and only then capturing it as a skill. Eliot starts by describing his pain point: turning transcripts into SEO-optimized blog posts that actually rank, not just AI-generated content. He invokes Claude in co-work mode and asks it to research best practices while incorporating his own expertise from past videos and teaching. Claude spends 20 minutes mapping out a multi-stage process, asking clarifying questions about scope and output format. Eliot gives his own expertise as transcripts about SEO methodology, which Claude folds in alongside fresh research. Once the SOP is drafted, he tests it by feeding it a transcript and requesting a blog post targeting the keyword 'Claude memory.' The output isn't perfect on the first pass, so he gives feedback: 'make titles more on-the-nose for SEO, put the keyword higher in H2s, add an FAQ section at the bottom.' Claude makes the edits. Eliot now has a polished output that reflects his exact standards. Only then, once he's locked in how the process should work, does he ask Claude to capture this as a skill. This method takes longer upfront but ensures the skill reflects your actual workflow, not a generic template. The house rules discovered during testing — em dashes, title case, numerals, on-the-nose titles — get baked into the skill instructions so they apply every time.
Method 3: Recording Your Screen to Show Claude Your Workflow
The third method uses Claude's record a skill feature, which works like creating a Loom video inside Claude. You click the plus button in co-work mode, select record a skill, and Claude begins capturing your screen, clicks, typing, and voice commentary. You then walk through your entire process step by step, narrating what you're doing. Eliot demonstrates this with an SEO reporting workflow: he opens Google Search Console, compares performance year-over-year, exports data, opens Google Analytics, runs live searches, creates a Google Doc with a structured report (good, bad, key insights, opportunities), and then drafts an email to the client in Gmail. Throughout, he narrates each step. Claude records 199 individual steps and screenshots matching them to his commentary. Once complete, Claude analyzes all that information and builds a skill that can repeat the process. The clever part is that Claude doesn't just replicate his exact clicks — it decides whether to use Claude in Chrome (the browser extension that lets Claude control your web browser), Gmail connectors, or other tools available to you. This method is best for workflows that involve clicking around websites, downloading data, or manual steps across multiple tools. It captures the full context of how you work without requiring you to describe it in words.
Managing, Editing, and Sharing Skills
Once a skill is built, you can manage it in the customize area. Skills can be toggled on and off if they're firing at the wrong time. You can edit trigger phrases or descriptions conversationally by asking Claude 'add five trigger phrases to this skill.' You can download any skill as a file and send it to someone else, or upload skills that others have created. In the customize section, you can see all your skills, click into any of them to view the full description, skill.md, and reference files, and even delete skills you no longer need. Trigger phrase management is important: if a skill fires when you don't want it to, you adjust the phrases to be more specific. Eliot added trigger phrases to his SEO blog writer skill in a second pass because they were initially missing. He conversationally edited the skill by saying 'add trigger phrases like turn this transcript into a blog post, help me write a blog post.' Claude updated it immediately. Skills also show differently in Claude co-work mode: on the right side of the interface, you can see which skills have been activated for your current task and click on them to inspect them. This transparency helps you understand which skill is handling which part of your workflow, especially valuable when running chains of multiple skills together.
Building an SEO Skill Team: From Individual Skills to Orchestrated Workflows
The course's central example builds a complete SEO department as a team of chained skills. An SEO specialist's workflow typically goes: audit the site, write a strategy based on findings, create content following the strategy, publish it, analyze results, report, and repeat. Rather than running each step manually, Eliot builds individual skills for each stage. The audit skill produces findings. The strategy skill receives those findings and writes a strategic plan. The content creator skill receives the strategy and writes blog posts. The publishing skill pushes the content live. A reporting skill pulls data from Search Console and Analytics. The email skill drafts client communications. Finally, an orchestrator skill — the manager — sits on top. You tell the orchestrator 'website traffic is down, what do we do?' and it knows which skills to call, in what order, and how to chain their outputs together. This level-4-and-5 architecture turns what was manual work into an automated production line. Each skill is a team member. The orchestrator is the line manager delegating work and coordinating handoffs. The business value comes not from any single skill but from the system: you describe the problem once, and the AI team breaks it into subtasks, executes them in sequence, and reports back. This is the vision Eliot is building toward throughout the course — shifting from 'I have a helper for individual tasks' to 'I have a department running end-to-end workflows.'